Legal protections for AI-generated work in China

DLA Piper lawyers Edward Chatterton and Wilson Lung explain a recent Chinese court judgment and review how luxury brands can protect copyright in AI-assisted work

The ruling means AI-assisted artworks can potentially be copyrighted in China Shutterstock

“How can we legally protect works we create using AI?” – this is a question intellectual property lawyers are frequently asked following the emergence of generative AI.

The Beijing Internet Court’s recent judgment in Li v Liu; Re Spring Breeze Has Brought Tenderness (case number (2023) Jing 0491 Min Chu No. 11279) provides important insight for luxury brands from the Chinese courts. In the decision, the court was willing to grant copyright protection over a work created by prompting an AI tool, an approach which appears more liberal than the position the authorities have taken in the US. However, the court’s decision very much depended on the facts of the case, demonstrating that the application of the legal principles can present complexities in practice.

The REAP method for protecting works involving AI in China

In light of this judgment, brands could benefit from applying the REAP method when using a generative AI model to create works in China. It’s an important way to enhance the likelihood of success in protecting works through copyright.

  1. Review the model’s terms of use carefully;
  2. Engage actively with the model using original prompts, and document your creation process, to show your contribution and control over the process;
  3. Add watermarks or copyright notices on your works upon publication; and
  4. Prepare yourself to reprise the skills and know-how you used with the model before the courts if asked.

Reviewing the case facts

The facts of the Li v Liu case are relatively simple. The plaintiff Li created several images through input prompts on Stable Diffusion, an open-source generative AI model created by Stability AI, including one named ‘春风送来了温柔’ (‘Spring Breeze Has Brought Tenderness’ in Chinese, which we’ll refer to as the image). Li then posted the images on Xiaohongshu (小红书), a popular social media platform in China. The defendant Liu downloaded the image, removed Li’s watermark on the image, and uploaded a copy (without the watermark) along with an article on Baijiahao (百家号), a reputable user-generated content platform in China. Li sued the defendant for infringement of the plaintiff’s right of authorship and his right of communication by information networks, both protected under China’s copyright law.

Key issue 1: copyright subsistence in the image

To qualify for copyright protection in China, a work must meet two conditions: (1) it must be an original intellectual creation in the realm of literature, art or science; and (2) it must be capable of being represented in a specific form. The court concluded that the image satisfied these conditions.

The court first held that the image, looking no different from an ordinary photograph or painting, falls within the realm of art, can be represented in a specific form, and qualifies as a work of fine art.

In addition, the court found that, by independently designing and refining visual elements and composition through several rounds of input prompts and parameter adjustments, and by making an artistic choice on his preferred dusk-lit close-up portrait image of a woman, the plaintiff engaged in a process of “intellectual creation” which resulted in an “original” work rather than a mechanical creation.

Key issue 2: ownership of the image’s copyright 

In general, copyright in a work created in China is owned by its author, who can be a natural person, a legal person or an unincorporated organisation. The court said that an AI model cannot itself be an author.

The AI model developer in this case, Stability AI, did not participate in creating the image. They solely acted as the producer of a creative tool. Moreover, the user licence expressly stated that Stability AI does not claim or retain any rights over the output content. As such, the court held that Li, as the author of the image, was the rightful copyright owner.

Key issue 3: the defendant’s acts of infringement

The court found the defendant’s unauthorised use of the image infringed the plaintiff’s right of communication by information networks, because this enabled the public to access the image at their preferred time and place. The court also found that the defendant’s removal of the watermark, which contained the plaintiff’s Xiaohongshu (小红书) account number that he chose as his signature, infringed the plaintiff’s right of authorship.

The court ordered the defendant to post a statement of apology on their Baijiahao (百家号) account for at least 24 hours to mitigate the impact of his infringing acts and to pay Li CNY 500 in damages.

Does it mean that China now recognises copyright subsistence in all AI-assisted intellectual creations?

No, the judgment does not go that far.

The plaintiff, as a natural person, made significant intellectual contributions through multiple input prompts, multiple parameter adjustments, and an artistic choice at the end. The judgment is explicit that the recognition of copyright in AI-assisted intellectual creations depends on the level of human involvement. If a user simply provides a basic instruction to a generative AI model, and is then satisfied with the first generated image, copyright may not exist. This is because copyright protects the expression of an idea rather than the idea itself.

There is a minimum threshold of human involvement required to meet the conditions of “intellectual creation” and “originality”. The plaintiff’s involvement exceeded the minimum, but neither the copyright law nor the judgment specifies the exact threshold required. The court commented in the judgment that whether AI-generated images reflect the author’s personalised expression depends on each individual case, and this cannot be generalised.

Currently, we can only say that the greater the user’s intellectual contribution when using a generative AI model, the more likely it is that copyright subsists in their AI-assisted intellectual creation.

If copyright subsists in an AI-assisted intellectual creation, does the AI user own that copyright? Can the user monetise it on their own terms?

In most cases, yes, but not always.

Tencent v Yingxun; Re Dreamwriter (case number (2019) Yue 0305 Min Chu No. 14010) is an early leading case on this point in China. Tencent is the developer of Dreamwriter, an intelligent writing assistant. In this case, the court recognised copyright subsistence in a specific article written by Dreamwriter for the following reasons: (1) Tencent’s creative team selected and developed the databases and algorithms used by Dreamwriter; (2) the team performed selection and arrangement processes; and (3) the algorithms were a technical implementation of the team’s creative work. In addition, the court ruled that Tencent, as a legal person, owns the copyright because they organised and supervised the creative team.

The judgment does not refer to the Dreamwriter case. As Stability AI waived any claim to copyright in the output content of Stable Diffusion, the court did not have any difficulty in finding the plaintiff, in his capacity as the author, to be the copyright owner.

However, what if the developers of a generative AI model made a greater contribution to the creation of the work? If the model: (i) was trained using a database consisting only of paintings and photographs selected by the developers; or (ii) was specifically programmed by the developers to generate images which always contain specific visual elements, should the user be the sole author of the generated image? To align the judgment with the Dreamwriter case, a possible solution is to regard the developers and the user as joint authors and, therefore, joint copyright owners of the image.

Putting aside any possible complications related to joint copyright ownership, the good news for users is that most of the popular generative AI models, including ChatGPT and Stable Diffusion, have terms of use where the model developers and providers waive any copyright they may have in a work created with the assistance of their models. In most cases, the user will own any copyright in the work created. As a result, they can sell or license the copyright in their work on their own terms.

Positive development

The judgment is a positive development. If it is not overturned on appeal, creators and businesses in China using generative AI models will have greater confidence that their intellectual creations are copyright-protected.

Edward Chatterton is partner and co-head of intellectual property and technology practice (Asia). He advises on all aspects of intellectual property law including trademarks, copyright, patents, advertising, brand protection, trade secrets, database rights and designs and can be reached at [email protected].

Wilson Lung is a senior associate in DLA Piper's intellectual property and technology team based in its Hong Kong office. He can be reached at [email protected]

Email your news and story ideas to: [email protected]